Hi Yves,
I have used lavaan in the past for SEM based
mediation analysis. Initially both my outcome and mediator variables
were continuous, where things are straightforward. Recently, I have been
working with a binary outcome, and I am confused about the validity of
this analysis. The basic method for mediation analysis is based on the
two regression equations (Baron and Kenny):
mediator = a*treatment + error
outcome = b*mediator + c*treatment + error
where
the indirect effect (a*b) can be estimated by the product of
coefficients method. Now if the outcome is binary, then we replace the
linear model above with logit or probit regression. In this case, work
by Imai et al. (see attached paper, page 773) argues that the product of
coefficients method is no longer valid (the estimate of a*b is not
asymptotically consistent) using the sequential ignorability
assumption. Does this mean that SEM in lavaan with a binary outcome, a
continuous mediator and a discrete treatment does not produce valid
results? I know that SEM has been extended to discrete endogenous
variables, but this extension has not specifically considered mediation
analysis. So is mediation analysis with a continuous mediator and a
binary outcome, estimating the indirect effect a*b, valid?
Thanks, Ina